Hacker News Reader

Comment by danso | Hacker News Reader

Love seeing these powerful, production libraries be as easy to install as `pip install blingfire`. They've even included a token-based classifier demo in Jupyter Notebook:

https://github.com/Microsoft/BlingFire/blob/master/doc/Bling...

View on HN

Yes, indeed, although Tensorflow and Pytorch have a long way to go before their GPU versions are equally easy to install. Their GPU versions are currently incredibly hard to install correctly, even on Linux. Some users just give up and use prebuilt containers that although good good for production code, can be hard to develop with.

Pytorch install without any problem for quite a long time (they have a whl with all libraries included). This can't be said about tensorflow unfortunately.

On Windows, in order to get Torch I need to use Anaconda.

Your name leaked from an identifier to a property of your account in my head and I actually thought this was a dead comment. Weird.

This is unrelated, but you should see the potential that this holds for blind people: https://arxiv.org/abs/1901.02527

In my own experience, I've found it much easier to install tensorflow with GPU support from scratch (including cuDNN...) than it is to install Spark/Hadoop.

And running a marathon is easier than doing Iron Man, but neither of them is a walk in the park.

This has been vastly improved fairly recently with the advent of the tensorflow-gpu meta package available via conda. See https://towardsdatascience.com/tensorflow-gpu-installation-m...

Reply on news.ycombinator.com